Why does AI that worked stop working?

The three drifts have different names and different fixes. Data drift: what comes in no longer looks like what used to (the market changes, a supplier's format, the type of customer), and the fix is retraining or readjusting with fresh data. Model drift: the provider updates or retires the model you were using and the same prompt returns something else, and the fix is pinning versions and testing every model change against a case set before adopting it. Environment drift: an API changes, a site you were reading, a platform's policy, and the fix is making the system fail loudly (block and alert) instead of degrading silently.

This house lives it in miniature: our pipeline runs with seven gates that block publication and several exist precisely because a drift reached us first (a format change that broke links, text reaching the reader raw). The rule that sums it up: if nobody looks at a metric every week, that system is already failing and you just do not know it yet. The pre-launch version of this problem (projects that die from choosing badly or not measuring) is its cousin, and the quarterly rescoring rhythm comes from the use-case matrix. The full framework, in AI systems.